DOI: 10.1515/orga-2017-0014
An Overview of Image Analysis Algorithms for License Plate
Recognition
Khalid ABOURA1, Rami AL-HMOUZ2
1 University of Dammam, College of Business Administration, Dammam, Saudi Arabia kaboura@ud.edu.sa
2 King Abdulaziz University, Department of Electrical and Computer Engineering, Jeddah, Saudi Arabia ralhmouz@kau.edu.sa
Background and purpose: We explore the problem of License Plate Recognition (LPR) to highlight a number of algorithms that can be used in image analysis problems. In management support systems using image object recog- nition, the intelligence resides in the statistical algorithms that can be used in various LPR steps. We describe a num- ber of solutions, from the initial thresholding step to localization and recognition of image elements. The objective of this paper is to present a number of probabilistic approaches in LPR steps, then combine these approaches together in one system. Most LPR approaches used deterministic models that are sensitive to many uncontrolled issues like illumination, distance of vehicles from camera, processing noise etc. The essence of our approaches resides in the statistical algorithms that can accurately localize and recognize license plate.
Design/Methodology/Approach: We introduce simple and inexpensive methods to solve relatively important prob- lems, using probabilistic approaches. In these approaches, we describe a number of statistical solutions, from the initial thresholding step to localization and recognition of image elements. In the localization step, we use frequency plate signals from the images which we analyze through the Discrete Fourier Transform. Also, a probabilistic model is adopted in the recognition of plate characters. Finally, we show how to combine results from bilingual license plates like Saudi Arabia plates.
Results: The algorithms provide the effectiveness for an ever-prevalent form of vehicles, building and properties management. The result shows the advantage of using the probabilistic approached in all LPR steps. The averaged classification rates when using local dataset reached 79.13%.
Conclusion: An improvement of recognition rate can be achieved when there are two source of information espe- cially of license plates that have two independent texts.
Keywords: Image Analysis; Probabilistic Modeling; Signal Processing; License Plate Recognition
1
Received: January 21, 2017; revised: March 30, 2017; accepted: May 28, 2017
1 Introduction
Image processing transforms information from the real world into image data represented by matrices. The ma- nipulation of these real valued matrices requires mathe- matical and statistical solutions in order to locate objects in the images. Optical character recognition (OCR) is the technology of identifying written or printed text that are
included in images using image-processing algorithms.
OCR is a well-explored problem. It has been used and commercialized in many applications that include LPR, libraries, banks, converting document into text searchable document in scanner and sorting mail in post offices. Some of our solutions apply in such contexts.
Many around the world contributed to research in li- cense plate recognition. LPR systems are in use in several countries, and Britain has an extensive ANPR (Automatic
Number Plate Recognition) CCTV network. Such achieve- ments took many years of research. However, the problem remains difficult to solve perfectly. For example, in many countries the law still requires the plate number of an in- fringing car to be recognized visually. This reliability issue keeps the LPR problem an active research topic. LPR uses video captured images to automatically identify vehicles through their license plate. LPR finds applications in theft prevention, parking lot management, hotels and prop- erty management, traffic laws enforcement, border con- trol etc. There are other methods to identify cars such as transponders, bar-coded labels and radio-frequency tags.
License plate reading remains however the way a vehicle is identified. LPR attempts to make the reading automatic by processing sets of images captured by cameras. There are three steps in the process; detecting a vehicle, trig- gering the captures of images related to that vehicle and treating those images for recognition of the characters in the license plate. The capture of the images, their transfer in digital form to a processor and the coordination of all tasks in a LPR system is a feasible engineering problem.
The processing of images for recognition is where research starts. LPR has three main parts; localization of license plate from image, segmentation of characters from local- ized license plate region and recognition of those charac- ters. These steps are performed automatically by software and require algorithms.
1.1 License Plate Localization
Plate localization is the first step in LPR. It aims to locate the license plate of the vehicle in an image. A variety of approaches have been proposed to localize a license plate in captured video images (Al-Hmouz and Aboura, 2014).
Some of the existing methods are morphological opera- tions, edge detection, corner detection, sliding concen- tric windows (Anagnostopoulos et al., 2006), fuzzy logic (Chang et al., 2006 ), Hough transform (Duc et al.,2005), neural networks (Kim et al. 2000), Fourier transform (Acosta, 2004), adaptive boosting (AdaBoost) algorithm (Dlagnekov, 2004). Al-Hmouz and Aboura (2014) intro- duce a new approach of plate localization using a statis- tical analysis of Discrete Fourier Transform of the plate signal. The plate signal is represented by five statistics:
strength of the signal, normalized maximum amplitude, frequency of maximum amplitude, frequency center and frequency spread. Combining with the color-based histo- gram thresholding (Aboura, 2008), the method achieves 97.27% accuracy using plate signals from binary images.
1.2 License Plate Segmentation
The second task in LPR is to extract the characters from the localized license plate region. The most common meth-
od is the projection method. First, the image is threshold- ed, reduced from a color or gray-scale image to a black and white (background/foreground) image. The projection method then counts the number of foreground pixels ver- tically and horizontally in the license plate area to sepa- rate and extract the characters. Other methods and variants include, thin window scanning, local vector quantization, scale shape analysis, Laplacian transform, Hough trans- form and Markov Random fields. The final task in LPR is the recognition of characters. An optical character rec- ognition algorithm is used to recognize characters in the image. A series of approaches were developed. The most common ones are the correlation-based template matching and neural networks. Other methods are feature based, use pattern mapping or are based on the Hausdorff distance.
Binary classifiers are used as well as the Hidden Mark- ov model and probabilistic modeling. All three steps in LPR rely on thresholding or binarization of the original image. A common approach is the well cited Otsu method (Otsu, 1979). The Otsu method performs rather poorly in LPR if applied to the whole image. The resulting binary image often doesn’t show the license plate characters as foreground. In Aboura (2008), this topic is discussed and a new thresholding method is introduced. It is used in the approaches described in this article.
1.3 License Plate Recognition
The final task in LPR is the recognition of characters. The extraction of characters results in a number of selected re- gions in the image. These regions contain the characters of the license plate and are processed for recognition by what is referred to as an Optical Character Recognition (OCR) algorithm. While OCR applies in LPR, it is a field that has a much wider stage of applications. OCR dates back to 1929 and is a method designed to translate automatically images of handwritten or typewritten text. There are of- the-shelves software packages that can translate a faxed page of text or an image that contains text. However, these OCR commercial packages do not yield good results in the case of LPR. A series of approaches were developed to rec- ognize license plate characters. The most common OCR approaches used in LPR are the correlation-based template matching and neural networks. Other methods are feature based, use pattern mapping or are based on the Hausdorff distance. Binary classifiers are also used as well as the Hidden Markov model (Aboura and Al-Hmouz, 2007).
Neural networks (NN) haven been applied successfully in many prediction, classification and recognition problems.
In LPR, they are used to localize the license plate in the image and to recognize the extracted characters of the plate. A Neural network is an artificial network made up of sets of interconnected nodes called neurons. In its simple form, the feed forward structure used in LPR, there is a set of input nodes, for example the features and attributes of
Figure 1: Binary image with foreground pixel =1 and background pixel = 0 the image being processed (Figure 1), that are connected
through a network of nodes, hidden layers, to a set of out- put nodes, the classes to which the image belongs.
Template Matching is the other approach most used in character recognition in LPR. It is a technique in image analysis for scanning an image template until part of it matches an image at hand. There are many variants in the application of template matching to character recognition.
In its simplest form, the image in its binary form (Figure 1) is compared with same size parts of the template image using a suitable metric. The metric can be the Euclidian distance or a correlation measure between the pixels of the image and the template. For example, the cross cor- relation, a statistical measure used by Horowitz (1957) and Pratt (1974) for image recognition, can be a metric for template matching. The template matching approach is combined with other methods in character recognition.
However, it remains a method based on the minimization of a distance between two images and can prove to be in- efficient in practice.
The objective of this paper is to present a number of probabilistic approaches in LPR steps, then combine these approaches together in one system. Most LPR approaches used deterministic models that are sensitive to many un- controlled issues like illumination, distance of vehicles from camera, processing noise etc. The essence of our ap- proaches reside in the statistical algorithms that can accu- rately localize and recognize license plate.
The structure of the paper is as follows, first we show the probabilistic approaches that investigate the localiza- tion and recognition phases in LPR systems. In section 3, we show the results of the presented approaches. Followed by discussion in section 4. Finally, the paper ends with conclusions in section 5.
2 Research Methods in Probabilistic Image Analysis
In this section, we review some of the statistical and signal processing methods that address the localization and char- acter recognition problems discussed in the introduction.
We also point to the global thresholding method used in the methodology.
2.1 Localization using the Discrete Fourier Transform signal
A variety of approaches have been proposed to localize a license plate in captured video images. Al-Hmouz and Aboura (2014) introduce a new approach of plate localiza- tion using a statistical analysis of Discrete Fourier Trans- form of the plate signal. The plate signal is represented by five statistics: strength of the signal, normalized maximum amplitude, frequency of maximum amplitude, frequency center and frequency spread. The idea is that license plate numbers show frequencies that lead themselves to a spec- tral analysis through the Fourier transform. Al-Hmouz and Aboura (2014) introduce a formal statistical analysis of the Fourier transform data from a systematic scanning of the image. The authors use the Hysteresis thresholding (Can- ny, 1986) over scanned regions of the image to obtain a clear signal. They improve the speed of the methodology and its accuracy by thresholding the whole image using the binarization approach of Aboura (2008). The behav- ior of the power spectrum of the scanned region shows a significant increase in magnitude at some frequencies for scanned regions that contain the license plate or parts of it. This is due to the periodicity in the signal generated by
Figure 2: Extracted character (a), cleaned (b), cropped (c) and normalized (d)
Figure 3: Historical means of projected foreground for the 36 characters the characters of the license plate. A set of 5000 images
was used as historical data. The images were taken at a parking lot entrance. Each image was visually inspected for the existence of a car in it. For the image signal, the five statistics were considered and used in a statistical model. A Bayesian analysis delivered high probability candidate re- gions in the image. To locate the license plate in one of the candidate regions, another probabilistic method was used that yielded the achieved accuracy.
2.2 Probabilistic Optical Character Recognition
The LPR optical character recognition is the problem where uncertainty prevails as to what class an input image belongs to. Given that it is a stochastic problem, one ex- pects probability answers, if one adheres to the principle that probability is the only coherent way to address uncer- tainty (Lindley, 1987). Despite some attempts, such as the probabilistic transition trees (Eichelberger and Najarian, 2006), the only noticeable probability based research di- rection is that of the probabilistic neural networks, for ex- ample (Anagnostopoulos, Anagnostopoulos, Loumos and Kayafas, 2006). The probabilistic neural network (PNN) was developed by Donald Specht (1988) and provides a
solution to classification problems using Bayesian clas- sifiers and the Parzen Estimators. It is a class of neural networks which combine statistical pattern recognition and feed-forward neural networks technology. It is char- acterized as having very fast training times and it produces outputs with Bayes posterior probabilities. PNNs are very effective for pattern recognition. However, the LPR char- acter recognition problem is a simple OCR problem and can be addressed by a full probabilistic approach such as the one we present. To solve the problem probabilistically is to treat features or statistics from the input image, using a probability model. The input character image to the OCR module is usually a binary image such as the one in Fig- ure 1. The historical data of extracted characters is made of two sets; training data and validation data. These are images of license plate characters that have been extract- ed from images of vehicles using the first two steps LPR.
They are binary images that have been cleaned, cropped and normalized (Figure 2). The characters are visually in- spected one by one and classified in the 36 possible classes {A,B,C, ..., X,Y,Z,0,1, ...,8,9}. Each set of characters is then split into a training set and a validation set.
Examples of statistics of an extracted character are the fill percentage and the projected foreground. The fill per- centage is the proportion of foreground pixels in the binary
image. If F is the matrix of the digitized binary image, then thefill percentage is ∑x ∑y F(x,y) / (Nx Ny), where Nx and Ny are the height and width of the image in pixels. The projected foreground is the normalized projection of the foreground on the xand y axes of the image ∑y F(x,y) / Ny, x = 1 ,..., Nx and ∑x F(x,y) / Nx, y = 1, ..., Ny respectively.
These statistics seem to provide distinguishing informa- tion, as seen in Figure 3 where the mean of these statis- tics is plotted within one standard deviation. We used the minimization of squared errors to attempt to recognize the characters using the normalized projection foreground:
Minimize(C=A,B,…,X,Y,Z,0,1,…,8,9)
∑x (∑y F(x,y) / Ny - µxC)2 + ∑y (∑x F(x,y) / Nx - µyC )2 where µxC and µyC, C=A,B,…,X,Y,Z,0,1,…,8,9, are the historical means from the training set. The approach failed to return a good recognition. In an attempt to remedy, the projected foreground was augmented with the distance of the pixels to the side of the image, thus incorporating the information about the locations of the foreground pixels.
The minimum squared errors method failed again, an indi- cation that these features of the image do not offer enough information to fully distinguish among the characters.
In general, let Z be the random variable that represents the statistical feature of a character image. For example, Z can be the fill percentage seen above. Note that Z need not, and often isn’t univariate. The probabilistic approach starts by building a probability model for that feature in the form of Prob(Z|C),C = A,B,...,9. For each image in the training set, the value of Z is computed. Data analysis tools are used, along with any engineering/prior knowledge to arrive at the probability model Prob(Z|C). Seen as a func- tion of the event C, that is the character is C, the probabil- ity model Prob(Z|C) is known as the likelihood function L(C) = Prob(Z|C). This likelihood function is at the heart of the probabilistic approach. If this model is built proper- ly and the statistical feature Z offers enough information about the character’s class, the probabilistic approach will be effective. Given a likelihood model, the probabilistic approach proceeds as follows. Let z be the value of Z for an image being analyzed. Then the probability that the char- acter is C, given the data z is
for C = A,B,...,9. Prob(C) is the model built with any prior knowledge about what C might be. It is called the prior distribution.
is a normalizing constant. Prob(C|Z=z) is the posterior dis- tribution of C. The solution is given in the selected charac-
ter Ĉ, mode of the posterior distribution,
If there is no prior knowledge as to what C might be, then the discrete uniform distribution is used where Prob(C) = 1/36 for C = A, B, ...,9. This probabilistic approach is often referred to as a Bayesian approach. A non-Bayesian approach would simply maximize the like- lihood function, ignoring the prior component, and select Ĉ such that
As one can see, both these statistical approaches rely on the likelihood function. These simple operations are at the heart of many probabilistic predictions, classifications and inferences. While seemingly simple in principle, their success depends on the proper selection of the random var- iable Z and the probability model Prob(Z|C).
In our search for a statistic Z, we arrived at the conclu- sion that the values of the pixels in the binary image hold all the information needed to recognize the character in an image that has been cleaned, cropped and normalized. We defined Z as the multidimensional vector of the values of all the pixels in the image. Each of these values is either 0 or 1, the input image being binary. For each pixel, we applied the Bernoulli probability model θi(Zi (1 - θi )1-Zi , Zi being the value of Z at pixel i. Making the assumption of conditional independence of the pixel values given an image, we construct the likelihood function
where |Z| is the cardinal, or vector size, of Z. To esti- mate the proportion θi for pixel i, a number of approaches are available. But given that the sizes of the historical sets are relatively large, in the order of 200 images per charac- ter C = A,B,...,X,Y,Z,0,1,...,8,9, the estimates converge to the average
where Nc is the size of the training set for character C, and xij = {0 or 1} is the value of pixel i for image j of the training set. This is done for each character C. For simplicity of exposition, we used θi when in fact it is a θi (C) that differs for each C. From a computational point of view, the assessment of the likelihood parameters is very simple. For each character C, all the image matrices of the training set are added, then divided by the size of the set Nc, automatically providing a matrix of estimates
This is a simple operation, inexpensive computational- ly, that replaces the training of a neural network. It needs to be done only once, and the estimates matrices are used subsequently to recognize the characters.
Figure 4 shows the matrix
Figure: 4 Likelihood image of character K
for character K. Each value of the pixel i of the image in Figure 4 is the estimate 0 ≤ θ̂i ≤ 1 for character K. It is the estimate of the parameter of the Bernoulli model for pixel i, for character K. In essence, this method builds a likelihood function value for each character C, that results from a matrix, where the values of 1 signify that the cor- responding pixels are always present in the foreground of C (in black in Figure 4), the values 0 mean that the corre- sponding pixels are always background in C (in white in Figure 4), and the values in between correspond to pixels i that are present in C with a positive probability θ̂. One observes that the K in the image is not a perfect one, due precisely to statistical variation. In addition, the colors of the image do not show all the nuances in the values. But the matrix
is computed with double precision accuracy and provides accurate estimates of the probabilities of the foreground existence.
Once the likelihood function is constructed, it is used to recognize characters in a simple operation. Let z be the realization of the statistic Z for a binary image that has re- ceived similar cleaning, cropping and resizing as have the images of the training set. Then
where
noting that the appropriate θ̂i correspond to a given C.
This posterior probability distribution ranks the charac- ters A,B,...,X,Y,Z,0,1,...,8,9 for their likelihood of being the character in the image being treated. This method pro- vides a full probabilistic approach. One introduces prior knowledge in the Bayesian analysis above in the form of Prob(C) for each character C . Such knowledge can vary from country to country for example, where the position of a character in the license plate may indicate whether it is a letter or a number, for example. If such knowledge does not exist, or is hard to embody in a formal model, then the characters are declared to be equally probable a priori, letting Prob(C) = 1/36 in this case. To further improve the speed and increase the accuracy, we note that maximizing the product of two bounded positive values is equivalent to maximizing their sum. Therefore we implemented the following
where f(C) is the score function defined by the sum. This method yielded excellent results.
2.3 Global Thresholding
Al-Hmouz and Aboura (2014) used a global thresholding method to improve significantly the performance of a sig- nal processing localization algorithm. This is in contrast to most methods that use a local thresholding approach in view of the fact that global thresholding is hard to per- form adequately. The global thresholding method of Abou- ra (2008) was effective in highlighting the signal of the license plate (Figure 5). A more effective global thresh- olding method experimented with consisted of the color- based method of Aboura (2008) coupled with a range of techniques including clustering. In a striking result, all but the characters of the license plate remained as foreground
in the thresholded image. Such an approach would elim- inate the need for a localization algorithm and facilitate greatly the recognition of the characters. In most cases, only some of the characters or part of remained as fore- ground. This still reduces considerably the localization of the license plate. This research was not completed and is a future direction of work.
3 Research Results
3.1 Object Localization in Image
In Al-Hmouz and Aboura (2014), a statistical localization approach of Discrete Fourier Transform and five statistics of the plate signal is used to extract the license plate in LPR system. The approach yielded 99% of license plates were among few candidate regions and the overall perfor- mance was 97.27% for local data set and 88.1% for public data set. The method is considered computationally com- plex compared to other methods in the literature. However, the processing time can be reduced to competitive values (150 ms) considering less number of sliding window lo- cations by increasing the step size of the sliding window and ignoring places where license plates do not exist such as image sides. Methods such as CCA and edge detection could be used in improving robustness after the plate can- didate regions have been determined. In the future, infor- mation about license plate images which are not localized will be studied through suitable models.
3.2 Probabilistic Optical Character Recognition
The LPR optical character recognition falls within the realm of pattern recognition. While such problems often require so- phisticated approaches due to the large number of patterns, known and unforeseen ones, making it hard to use parsimo- nious probability models, the recognition problem in LPR is much simpler. We introduced a simple, and very inexpensive method to solve a relatively important problem, using a full probabilistic approach. As most methods in this category, the approach failed to distinguish as such fully between some characters like 2 and Z, 5 and S, 1 and I, B and 8, and O,0,D and Q. However, using the same logic and applying it exclusively to parts of the image, we reach a 97% relia- bility. In Figure 6, we show how 2 and Z are distinguished by focusing the probabilistic OCR algorithm in the areas of the boxes. The likelihood of the pixels in those areas has already been computed, and it is applied to the image that was initially recognized either as a 2 or a Z. Similarly, Figure 7 shows how the remaining confusions are dealt with. This added step is a refinement equivalent to mov- ing a magnifier glass, in the form of the probabilistic OCR algorithm, to parts of the image. Although the likelihood function is the same, the scoring changes and the choice is between two, or four, rather than 36 candidates.
The algorithm was tested on a large validation set, of the order of a couple of hundreds of images per character.
Table I shows the reliability per character, summing up to an overall 0.9705 reliability. One could possibly conduct comparison tests with other methods. But as it is often the case in LPR problems, the scenarios and setups that pro- vide sources of data vary dramatically, making it hard to compare universally. We used two sources of data; images
Figure 5: Image thresholded using global thresholding method
taken live from a LPR system setup in a parking lot en- trance, and images taken from the entrance of the parking lot of the university. The lighting conditions and the posi- tioning of the cameras differed between the two locations.
In addition, two slightly different character segmentation methods were used in each case. The tests conducted per- formed well on both types of data mixed together, as indi- cated in table I.
3.3 Global Thresholding
A global thresholding method for the automatic license plate recognition is introduced that makes use of the colour infor- mation in the image through a probabilistic model. LPR still finds it difficult to apply global binarisation whereby a sin- gle threshold value for the entire image is calculated. Local thresholding methods are preferred, but their efficiency comes at a cost. Reported work is often focused on one of the three LPR steps, and thresholding is assumed effective. Threshold- ing a range of images in an automatic fashion is not a trivial task. Aboura (2008) introduces the concept of a statistically determined global threshold for LPR images that is shown to be effective on a large data set. Although the approach is en- veloped in a linear regression model, it is the way information from the image is selected and introduced in the explanatory variables that leads to the reliable results. The reported meth- odology proposes a model that can be used in conjunction with other methods to result in automated systems.
4 Discussion
All the previous approaches adopted probabilistic models in LPR steps. Each approach improved the recognition rate for the plate with regard to localization recognition rate and character recognition rate. Using both approaches in LPR system yielded an average recognition rate that reaches up to 97.13%. As mentioned earlier, the confusion between some characters (refer to figure 6 and 7) might have the major drawback of this algorithms. However, this problem will not have an effect on bilingual plates in which two independent texts are appeared in the plates.
In some non-English speaking countries like Saudi Arabia, license plate appears in English and Arabic text (Figure 8). Arabic letters are associated with their corre- sponding English letters. This can be considered as two source of information available for character recognition.
In this case, the probabilities of Arabic and English char- acters can be combined in character recognition phase. The posterior of each Arabic characters is fused with its associ- ated English characters. The prevailing character probabil- ities after fusion will be considered as the final recognition.
The flow chart of license plate recognition for bilingual plates is show in Figure 9.
Figure 6: Probabilistic OCR applied to eliminate Confusions
Figure 7: Refinement of the Probabilistic OCR
Table 1: OCR Reliability Results
Character Reliability Character Reliability
A 0.9888 S 0.9385
B 0.9259 T 0.9828
C 0.9841 U 1.0000
D 0.9552 V 0.9306
E 0.9359 W 0.9898
F 1.0000 X 1.0000
G 0.9608 Y 1.0000
H 1.0000 Z 0.9722
I 0.9500 0 0.8636
J 1.0000 1 0.8864
K 1.0000 2 1.0000
L 1.0000 3 0.9649
M 1.0000 4 0.9899
N 0.9529 5 0.9623
O 0.9630 6 0.9756
Figure 8: Saudi Arabia license plate examples
Figure 9: Bilingual plate (Arabic English) processing steps
5 Conclusions
In this paper, we described a number of solutions that can be used in recognition systems for decision support. These algorithms are effective alongside the application of global thresholding and other techniques that uses frequency of signals from the images by means of the Discrete Fourier Transform. This paper presented a probabilistic character recognition approach. The likelihood functions were formed using the Bernoulli model for each pixel value and the posterior probabilities of characters were obtained by fusing the probabilities of pixel value in each location. In making the assumption of independence between pixels of the image in the construction of the likelihood function, we in fact constructed a Naïve Bayes solution. The results showed an impressive recognition rate of average (97.13%) when using the previous approaches. Also an improvement of recognition rate can be achieved when there two source of information especially of license plates that have two independent texts. For further study, fusion process at pixel level for bilingual plates can investigated.
References
Aboura, K. (2008). Automatic Thresholding of li- cense plate. International Journal of Automation and Control, 2(2-3), pp. 213-231. DOI: 10.1504/
IJAAC.2008.022178.
Aboura, K. & R. Al-Hmouz, R. (2007). Probabilistic li- cense plate optical character recognition. School of Computing and Communication, University of Tech- nology Sydney. (Technical Report K-OCR-2-07) Acosta B.D. (2004) Experiments in image segmentation
for automatic US license plate recognition. Master’s thesis, Virginia Polytechnic Institute and State Univer- sity.
Al-Hmouz, R., & Aboura, K. (2014). License plate local- ization using a statistical analysis of discrete fourier transform signal. Computers and Electrical Engineer- ing, 40(3), pp. 982–992, http://dx.doi.org/10.1016/j.
compeleceng.2014.01.001
Anagnostopoulos, C., Kayafas, E., & Loumos, V. (2000).
Digital image processing and neural networks for ve- hicle license plate identification. Journal of Electrical Engineering, 1(2), pp. 2-7.
Anagnostopoulos, C., Anagnostopoulos, I., Loumos, V.,
& and Kayafas, E. (2006). A license plate recognition algorithm for intelligent transportation system applica- tions. IEEE Transactions on Intelligent Transportation Systems, 7(3), pp. 377-392, http://dx.doi.org/10.1109/
TITS.2006.880641
Canny, J. (1986). A computational approach to edge de- tection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 8 (6), pp. 679-698, http://dx.doi.
org/10.1109/TPAMI.1986.4767851
Chang S.L, Chen. LS, Chung Y.C, Chen S.W. (2004).
Automatic license plate recognition. IEEE Trans Intell Transport Syst, (1), pp. 42–53, http://dx.doi.
org/10.1109/TITS.2004.825086
Dlagnekov L. (2004). License plate detection using Ada- Boost. Comput. Sci. Eng. Dept., Univ. California, San Diego, CA; March 2004. Retrieved from http://www.
cse.ucsd.edu/classes/fa04/cse252c/projects/louka.pdf Duc DT, Hong T.L, Vinh P.T, Viet H.N. (2005). Building
an automatic vehicle license plate recognition system.
In Proc. Intl. Conf. in Computer Science – RIVF’05 February 21-24, 2005, Can Tho, Vietnam, pp. 59–63.
Eichelberger, C.N., & Najarian, K. (2006). Simulating protein computing: character recognition via probabil- istic transition trees. IEEE International Conference on Granular Computing, pp. 101-105, http://dx.doi.
org/10.1109/GRC.2006.1635766
Horowitz, M. (1957). Efficient use of a picture correlator.
Journal of the Optical Society of America, 47(4), pp.
327-327, http://dx.doi.org/10.1364/JOSA.47.000327 Lindley, D.V. (1987). The probability approach to the
treatment of uncertainty in artificial intelligence and expert systems. Statistical Science, 2(1), pp. 17-24, http://dx.doi.org/10.1214/ss/1177013427
Kim KK, Kim KI, Kim JB, Kim HJ. (2000). Learn- ing-based approach for license plate recognition. In:
Proc. 2000 IEEE signal processing society workshop, Sydney, Australia, pp. 614–23.
Otsu, N. (1979). A threshold selection method from gray-level histogram. IEEE Transactions on System, Man and Cybernetics, 9(1), pp. 62-66, http://dx.doi.
org/10.1109/TSMC.1979.4310076
Pratt, W.K. (1974). Correlation techniques of image regis- tration. IEEE Transactions on Aerospace and Electron- ic Systems, 10, pp. 353-358, http://dx.doi.org/10.1109/
TAES.1974.307828
Specht, D.F. (1988). Probabilistic neural networks for classification, mapping or associative memory. IEEE International Conference on Neural Networks, 1, pp.
525-532, http://dx.doi.org/10.1109/ICNN.1988.23887
Khalid Aboura, PhD, teaches quantitative methods at the College of Business Administration, University of Dammam, Saudi Arabia. Khalid Aboura spent several years involved in academic research at the George Washington University, Washington D.C, U.S.A. He worked as a Research Scientist at the Commonwealth Scientific and Industrial Research Organization of Australia and conducted research at the University of Technology Sydney. Khalid Aboura was a Scientist at the Kuang-Chi Institute of Advanced Technology of Shenzhen, China.
Rami Al-Hmouz received his PhD degree in computer engineering from the University of Technology, Sydney, Australia, in 2008. Currently, he is an Assistant Profes- sor in Department of Electrical and Computer Engineer- ing at King Abdulaziz University in Saudi Arabia. His research interests are in machine learning, fuzzy logic, granular computing, and computer vision.
Pregled algoritmov za analizo slike za prepoznavanje registrske tablice
Ozadje in namen: V članku raziskujemo problem prepoznavanja registrskih tablic (LPR), in podamo pregled števil- nih algoritmov, ki jih lahko uporabimo pri problemih analize slik. V sistemih za podporo vodenju, ki uporabljajo za prepoznavanje slikovnih objektov, je inteligenca vgrajena v statistične algoritme, ki jih je mogoče uporabiti v različnih korakih razpoznavanja. Opisujemo več rešitev, od začetnega koraka do lokalizacije in prepoznavanja slikovnih el- ementov. Cilj tega prispevka je predstaviti več verjetnostnih pristopov v korakih razpoznavanja, nato pa združiti te pristope v en sistem. Večina pristopov uporablja deterministične modele, ki so občutljivi na številne nenadzorovane vplive, kot so osvetlitev, razdalja vozila do kamere, šum pri procesiranju itd. Bistvo naših pristopov je v statističnih algoritmih, ki lahko natančno lokalizirajo in prepoznajo registrsko tablico.
Oblikovanje / metodologija / pristop: Predstavimo enostavne in poceni metode za reševanje relativno pomemb- nih problemov z uporabo verjetnostnih pristopov. Pri teh pristopih opisujemo številne statistične rešitve od stopnje začetnega praga do lokalizacije in prepoznavanja slikovnih elementov. V koraku lokalizacije uporabljamo frekvenčne signale iz slik registrskih tablic, ki jih analiziramo z uporabo diskretne Fourier-jeve transformacije. Pri prepoznavanju znakov na tablicah smo uporabili tudi verjetnostni model. Na koncu prikazujemo, kako združiti rezultate iz dvojezičnih tablic, kot so na primer tablice Saudove Arabije.
Rezultati: Algoritmi so učinkoviti pri razpoznavanju znakov na vozilih, v stavbah in drugod. Rezultat kaže prednost uporabe verjetnostnega pristopa v vseh korakih razpoznavanja registrskih tablic. Povprečne stopnje uspešnega raz- poznavanja pri uporabi lokalnega nabora podatkov so dosegle 79,13%.
Zaključek: Izboljšanje stopnje razpoznavanja je mogoče doseči, če obstajata dva vira informacij, še posebej na registrskih tablicah, na katerih sta dve neodvisni besedili.
Ključne besede: analiza slike; verjetnostno modeliranje; obdelava signalov; prepoznavanje registrske tablice
